Black Forest Labs officially released the FLUX.2 series, launching four models: pro, flex, dev, and klein (upcoming). It also open-sourced the 32B parameter FLUX.2 [dev] weights and code, claiming it as the strongest open-source text-to-image model with multi-image reference and 4MP editing tools.

image.png

Core Upgrades  

- 10 Image References: Upload up to 10 style/composition examples at once, generating consistency >95%, supporting pose, lighting, and color locking  

- 4MP Editing: Local re-drawing, watermark removal, and background changes with input and mask, resolution increased to 4MP, PSD layer export is on the roadmap  

- Text Rendering: Accurate rate of 92% for complex layouts, infographics, and UI interface texts, exceeding DALL·E3 by about 10 percentage points  

- Realistic Logic: Error rate for spatial relationships, lighting physics, and world knowledge decreased by 37%, making meme and joke image generation more "reliable"

image.png

Four Model Tiers  

- pro: Commercial flagship, speed improved by 2 times, cost reduced by 30%, officially positioned as a "closed-source model alternative"  

- flex: Adjustable steps/ guidance coefficient, suitable for developers for secondary optimization  

- dev: 32B open-source weights, text-to-image + multi-image editing in one, Apache 2.0 license  

- klein: A lightweight distilled version to be released, reducing parameters by 50% with the same performance, targeting edge devices

Developer-Friendly  

- New VAE Autoencoder: Potential space compression rate improved by 18%, saving about 15% of training and inference GPU memory  

- Sample Code: PyTorch, Diffusers, and ComfyUI plugins are now available, and an introductory script is already on GitHub  

- Online Demo: huggingface.co/spaces/BlackForestLabs/FLUX-2-dev allows no-code experience

Industry Impact